EPPS: Advanced Polyp Segmentation via Edge Information Injection and Selective Feature Decoupling

TL;DR

This paper tackles polyp segmentation in colonoscopy images, where edge ambiguity and background noise hinder accurate masks. It presents EPPS, an encoder–decoder network augmented with an Edge Mapping Engine (EME) to extract edge cues, an Edge Information Injector (EII) to weave edge information into Decoder features via serial Channel and Spatial CBAM attention, and a Selective Feature Decoupler (SFD) that splits encoder features into significant and unimportant components using Mutual Information Neural Estimation (MINE) to minimize redundancy. The model is trained end-to-end with a joint loss $loss_{joint}=loss_{mask}+\alpha loss_{edge}+\beta loss_{MI}$, where $loss_{mask}$ combines Dice and BCE, and $loss_{edge}$ comes from edge predictions $edge_{pred}$ against $edge_{gt}$ generated by edge operators on ground-truth masks. A Feature Fusion Module (FFM) aggregates multi-scale Decoder outputs to produce precise polyp masks, and extensive experiments across three benchmarks show state-of-the-art performance, with ablations validating the contributions of EME, EII, SFD, and FFM and their synergistic effect on segmentation accuracy.

Abstract

Accurate segmentation of polyps in colonoscopy images is essential for early-stage diagnosis and management of colorectal cancer. Despite advancements in deep learning for polyp segmentation, enduring limitations persist. The edges of polyps are typically ambiguous, making them difficult to discern from the background, and the model performance is often compromised by the influence of irrelevant or unimportant features. To alleviate these challenges, we propose a novel model named Edge-Prioritized Polyp Segmentation (EPPS). Specifically, we incorporate an Edge Mapping Engine (EME) aimed at accurately extracting the edges of polyps. Subsequently, an Edge Information Injector (EII) is devised to augment the mask prediction by injecting the captured edge information into Decoder blocks. Furthermore, we introduce a component called Selective Feature Decoupler (SFD) to suppress the influence of noise and extraneous features on the model. Extensive experiments on 3 widely used polyp segmentation benchmarks demonstrate the superior performance of our method compared with other state-of-the-art approaches.

Figures (5)

• Figure 1: a)The overall architecture of EPPS. b)The Edge Mapping Engine. c)The Decoder block with Edge Information Injector. In this figure, CBR represents Convolution, Batch Normalization and ReLU. CAM and SAM stand for Channel Attention Module and Spatial Attention Module, respectively.
• Figure 2: a) Structure of the Selective Feature Decoupler (SFD). b) Structure of the Feature Fusion Module (FFM).
• Figure 3: Additional results on CVC-ClinicDB and Kvasir-SEG datasets. In e), for the significant feature maps ($s_i$), we outline the edge-related features learned by SFD with green boxes. In f), for the unimportant features ($u_i$), we outline the bright areas with red boxes, primarily located in areas of reflection and foreign objects, which are ineffective or could even mislead the model.
• Figure 4: The t-SNE visualization of the two features outputted by SFD ($s_i$ and $u_i$) with and without mutual information constraint. Red indicates $s_i$, while blue indicates $u_i$.
• Figure 5: Visualization of the model's mIoU scores on the CVC-ClinicDB dataset under different combinations of $\alpha$ and $\beta$, using a 3D surface plot.